Anthropic has put the activity of Claude-powered agents into an open industry format, and the format matters more than the feature. Agents built on Claude Managed Agents, the Messages API or the Claude SDKs can now send traces to any monitoring tool that speaks OpenTelemetry, the open standard for software telemetry. A trace is the step-by-step record of what a program did. According to Anthropic’s announcement on its Claude Resources site, the traces use the GenAI semantic conventions from OpenTelemetry, a shared naming scheme for AI calls. The upshot is that Claude’s data and other providers’ data can land on the same dashboards.

That is the gap between a vendor console and telemetry you own. A team running agents on two providers can compare them side by side, with the same charts and filters, instead of reading two incompatible consoles and reconciling them by hand. One caveat: the post names no rival provider whose output has been tested next to Claude’s. The comparison works only if the other vendor emits the same conventions.

Working out why an agent did something strange is one of the harder parts of running one. That is our observation, not Anthropic’s, and it echoes two other stories today: one asks whether adding agents to a late project repeats the 1975 man-month mistake, and another covers Google code that rewrites an agent’s own instructions. Traces are the plumbing for that kind of checking. Anthropic’s own advice is narrower: look at an agent’s traces before giving it more scope, and expand that scope only once it has earned the trust.

Tool use, model calls and whole agent turns each show up as their own trace, carrying latency, token counts and IDs. The post also lists calls to MCP servers (the connectors that let agents reach outside tools) among the steps worth tracing. It does not mention errors as a separate field, though debugging failed steps is one of its three named uses.

Those uses map to different tools. For debugging beside logs and metrics, Anthropic points to Splunk, Sentry, Honeycomb and Grafana Cloud. For turning real traces into tests and quality checks, it lists PostHog, New Relic, LangSmith, Langfuse, Datadog, Braintrust and Arize AI. For organisation-wide tracking, such as token use by team and unusual activity, it names Databricks Unity Gateway. Other OpenTelemetry tools should work too, provided they offer a standard OTLP endpoint (the receiving address most of them expose).

Ramp, a customer quoted in the post, offers the one number. It says that scoring its agents’ Langfuse traces with an internal system called Reflect has cut token usage by 10 to 20 percent on some agents and made sessions 30 percent faster. That is Ramp’s own account, and it does not isolate the effect of Claude’s new export.

The release is a public beta. With Managed Agents and the Messages API you first set up a destination inside the Claude Console, whereas the Claude SDKs send their traces without any setup. The post does not say where SDK traces go by default.

Two things are missing. Anthropic does not say whether the feature costs anything. More important, it does not say whether traces include prompt and response content or only the metadata it lists. For a company whose agents handle customer or financial data, that answer decides whether a third-party dashboard is allowed to receive these traces at all, so read the documentation before pointing anything at production.

Based on Anthropic’s Claude Resources post of 14 October 2026